Representation-level explanation of dialectal robustness

Determine, through representation-level analyses such as diagnostic classifiers for syllable boundaries, whether syllable-aware tokenization produces encoder representations that are more stable across dialectal phonetic variation.

Background

The results show that multisyllable- and syllable-aware tokenizers can improve word error rate on dialectal RVG 1 speech, and the paper hypothesizes that syllable-level CTC targets encourage representations aligned with syllable structure. This proposed mechanism is not directly tested by the reported WER, CER, or confusion analyses.

The authors identify representation-level diagnostics, particularly classifiers for syllable boundaries, as the direct way to test whether syllable-aware tokenization actually induces more stable representations under phonetic variation.